Phi-2-MATH

This is a colab notebook for finetuning Microsoft's Phi-2-3B LLM for solving mathematical word problems using QLoRA, Uploading adapters to 🤗 Hub, Merging the adapters and then uploading it on 🤗 repo. The notebook also contains code for inferencing it directly from my repo.

• Link to my repo: https://huggingface.co/ZappY-AI/phi2-math-orca

• Link to the dataset: https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k

• Base model: https://huggingface.co/microsoft/phi-2

• The model was loaded and trained in 4-bit quantization and float16 tensors for the sake of efficiency.

• The model was chosen because of its small size and less trainable parameters and extremely good performance as it can outperform models that are 5x-10x times bigger than itself. Take a look at these eval metrics:

• The model was trained for 500 steps on a T4 colab pro GPU (16 GB VRAM) for about 2.5 hours on a subset(20%) of the original dataset using TRL's SFT Trainer.
• The training loss obtained at the final step was 0.556700
• The following is the PEFT Config for this training notebook:

  peft_config = LoraConfig(
        lora_alpha=16,
        lora_dropout=0.05,
        r=16,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules= ["Wqkv", "out_proj"])
  

  • The following are the training metrics: